GPT-5.6 and the Sol Tier: OpenAI's Answer to Claude 5

OpenAI shipped GPT-5.6 on July 21, thirteen days after Claude Opus 5. The counter-punch timing is now so consistent that you can practically set a calendar by it. Anthropic ships, OpenAI ships two weeks later, and the announcement carefully avoids naming the competitor while responding to it point by point.

GPT-5.6: Same Price, Better Model

The base model holds at $5 per million input and $30 per million output, unchanged from GPT-5.5. Improvements are in instruction following, tool-call accuracy and refusal calibration. HumanEval hits 95.4, SWE-bench Verified 75.8, MMLU 93.1.

That puts it fourth on our TC Score leaderboard at 95.2, behind Claude Fable 5, Opus 5 and Sol. Comfortably competitive, not dominant. The streaming stack is noticeably faster, which matters more for interactive products than the benchmark deltas do.

Sol Is the Interesting Part

Sol is a configuration of GPT-5.6 that spends substantially more reasoning tokens per request by default. It posts 89.6 on GPQA, which is the best publicly reported score on that benchmark from anyone, and 97.8 on MATH.

It costs $12 per million input and $72 per million output. That is roughly 2.4x the base model, and because it also generates far more reasoning tokens, real-world cost per completed task is more like five to eight times higher.

This is not a general-purpose model. It is a scalpel. Use it for problems where a wrong answer is expensive and a slow answer is fine: hard mathematics, scientific analysis, complex architectural decisions, security review. Using it as a chat backend is setting money on fire.

Estimate what it would actually cost you on the price comparison tool before you wire it into anything.

The Mini and Codex Tiers

GPT-5.6 Mini at $0.60 per million input keeps the same tokenizer and tool interface as the full model. It scores 87.4 on MMLU and 55.3 on SWE-bench Verified. For high-volume classification, summarisation and sub-agent work it is strong value, and being interface-compatible with the full model makes routing between them trivial.

GPT-5.6 Codex is the coding-tuned build for the Codex CLI, IDE extension and cloud agents. Same price as base GPT-5.6, optimised for long autonomous sessions and patch generation rather than open-ended conversation. If you are doing agentic coding on OpenAI, this rather than the base model is the correct default.

How It Stacks Against Claude 5

Roughly: OpenAI wins on hard reasoning, Anthropic wins on long-horizon agentic work, and the gap in both directions is smaller than the marketing on either side suggests.

Concretely, Sol beats everything on GPQA and MATH. Claude Fable 5 and Opus 5 beat everything on SWE-bench Verified and hold up better across day-long agent runs. GPT-5.6 base and Claude Sonnet 5 are close enough on general work that the deciding factor should be price and ecosystem rather than capability. Sonnet 5 is cheaper at $3 per million input.

What To Actually Do

Already on OpenAI? Move to GPT-5.6, it is free capability. Use Codex for agentic coding.
Doing hard maths or science? Sol is worth benchmarking, budget permitting.
Doing long autonomous coding runs? Claude Opus 5 still has the edge.
High volume, quality-tolerant? GPT-5.6 Mini or Gemini 3.6 Flash, and route the hard cases up.

Full specs on the GPT-5.6 and GPT-5.6 Sol pages.